---
id: 20260815-T0-11
title: "别分类，直接让大模型\"幻觉\"出标签"
title_en: "Skip Classification, Let the LLM Hallucinate Tags"
url: https://ai.daily.yangsir.net/daily/20260815-T0-11
issue_date: 2026-08-15
publish_date: 2026-08-14T21:54:35.000Z
category: insight
source_name: "Simon Willison"
source_url: https://simonwillison.net/2026/Aug/14/dont-classify-hallucinate/
---

# 别分类，直接让大模型"幻觉"出标签

Simon Willison 介绍了 Doug Turnbull 提出的博客标签方案：不再让 LLM 从 1856 个标签中挑选匹配项，而是直接要求模型"幻觉"出合适的标签。这种方法避免了将大量标签一次性输入给模型导致的上下文过载问题。模型在"幻觉"模式下似乎能更专注于内容本身，生成更贴合文章主题的标签，即使这些标签并不在预设列表中。这一思路为处理大规模标签分类任务提供了新的启发。

## English Version

**Skip Classification, Let the LLM Hallucinate Tags**

Simon Willison highlights Doug Turnbull's approach to blog tagging: instead of asking an LLM to pick from 1,856 existing tags, simply tell the model to 'hallucinate' appropriate tags. This avoids overwhelming the model with too many choices in context. In 'hallucination' mode, the model focuses more on the content itself, generating tags that fit the article's theme, even if they're not in the predefined list. This offers a fresh approach to large-scale classification tasks.

---

**来源**：[Simon Willison](https://simonwillison.net/2026/Aug/14/dont-classify-hallucinate/)

**详情页**：https://ai.daily.yangsir.net/daily/20260815-T0-11

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*